Evaluation of a GPU-Based Large-Eddy Simulation for Dispersion in the Atmospheric Boundary Layer

Schalkwijk et al. (2015) recently conducted large-eddy simulations (LESs) of meteorological fields on a very large horizontal domain, 400 km X 400 km or the size of the Netherlands, with 100-m resolution. Their purpose was to simulate several weather scenarios -- fair weather cumulus clouds, cloud streets, and severe thunderstorm development -- in a timely manner, where computational "timeliness" was achieved using graphical processing units (GPUs). GPUs have undergone significant recent development with support from the commercial gaming industry and hold a key advantage in that they permit massively parallel computations with a large time savings. The new model GALES, a GPU-resident Atmospheric LES (Schalkwijk et al., 2012) was based on the earlier DALES, Dutch Atmospheric LES (Heus et al., 2010). For their exploratory runs, Schalkwijk et al. (2015) found that the ratio of computing-to-real time was about 4:1, an impressive achievement for such large calculations.

In this study, we apply GALES to dispersion in the convective boundary layer (CBL) for a much smaller domain than above, 5 km X 5 km X 2km (x,y,z), where the grid sizes are 52 m and 21 m in the horizontal (x,y) and vertical (z) directions. Our purpose is to evaluate the GPU-LES performance in driving a Lagrangian particle dispersion model (LPDM) in simulations of dispersion from several source heights in the CBL. The LPDM and CBL conditions are the same as used in Weil et al. (2004, 2012), which serves as the reference case for judging the model performance; the CBL has a height of 1000 m, a mean wind speed of 3 m/s, and a convective velocity scale of 2 m/s. We found that the GPU-LES-LPDM combination performed as well as the LPDM-LES in Weil et al. and was in agreement with the Willis and Deardorff (1976) laboratory experiments and the CONDORS field observations (Briggs, 1993; Weil et al., 2012). This initial pilot study was done with the LPDM running on a central processing unit (CPU), i.e., external to the GPU which produced the velocity fields, and thus the overall run time was not greatly reduced compared to the earlier simulations. The next step in the development is to put the LPDM inside the GPU, which should substantially lower the run time and permit the model to run faster than "real time."